Command-Line Interface

The CLI (tvc) provides command-line tools for model building, benchmarking, and export.

Installation

The CLI is installed automatically with the package:

pip install torchvision-customizer

# Verify installation
tvc --version

Commands

build

Build a model from a YAML recipe.

tvc build --yaml model.yaml [options]

Options:

--yaml, -y: Recipe YAML file (required)
--output, -o: Output path for saved model
--num-classes: Override number of classes
--format: Output format (pt, onnx, torchscript)

Examples:

# Build and save model
tvc build --yaml model.yaml --output model.pt

# Build with custom classes
tvc build --yaml model.yaml --num-classes 10 --output model.pt

# Build and export to TorchScript
tvc build --yaml model.yaml --format torchscript --output model.ts

benchmark

Benchmark model performance.

tvc benchmark --yaml model.yaml [options]

Options:

--yaml, -y: Recipe YAML file (required)
--batch-size: Batch size (default: 16)
--input-size: Input image size (default: 224)
--device: Device (cpu, cuda, mps)
--warmup: Warmup iterations (default: 5)
--iterations: Benchmark iterations (default: 50)

Example:

tvc benchmark --yaml model.yaml --device cuda --batch-size 32

# Output:
# Results:
#   Mean latency:    12.34 ms
#   Throughput:      2596.2 images/sec
#   Peak GPU memory: 1024.5 MB

validate

Validate a recipe file.

tvc validate --yaml model.yaml [--strict]

Options:

--yaml, -y: Recipe YAML file (required)
--strict: Strict validation mode

Example:

tvc validate --yaml model.yaml
# Recipe is valid!

tvc validate --yaml broken.yaml
# Validation Error: stages must be a list

export

Export model to different formats.

tvc export --yaml model.yaml --output model.onnx [options]

Options:

--yaml, -y: Recipe YAML file (required)
--output, -o: Output file path (required)
--format: Export format (onnx, torchscript)
--input-size: Input image size (default: 224)
--opset: ONNX opset version (default: 11)

Examples:

# Export to ONNX
tvc export --yaml model.yaml --format onnx --output model.onnx

# Export with custom input size
tvc export --yaml model.yaml --format onnx --input-size 384 --output model.onnx

# Export to TorchScript
tvc export --yaml model.yaml --format torchscript --output model.pt

list-backbones

List all supported torchvision backbones.

tvc list-backbones [--filter PATTERN]

Options:

--filter: Filter by name pattern

Example:

tvc list-backbones
# Resnet:
#   - resnet18
#   - resnet34
#   - resnet50
#   ...

tvc list-backbones --filter efficient
# Efficientnet:
#   - efficientnet_b0
#   - efficientnet_b1
#   ...

list-blocks

List all available building blocks.

tvc list-blocks [--category CATEGORY]

Options:

--category: Filter by category (block, attention, regularization, etc.)

Example:

tvc list-blocks
# Block:
#   - conv
#   - residual
#   - mbconv
# Attention:
#   - se
#   - cbam_block
#   - eca

tvc list-blocks --category attention
# - cbam_block
# - channel_attention
# - eca
# - se

create-recipe

Create a recipe from a template.

tvc create-recipe --template NAME --output FILE [options]

Options:

--template, -t: Template name (required)
--output, -o: Output YAML file (required)
--num-classes: Number of classes

Available templates:

resnet_base: Basic ResNet architecture
efficientnet_base: EfficientNet-style architecture
hybrid_resnet_se: ResNet with SE attention (hybrid)

Example:

tvc create-recipe --template hybrid_resnet_se --output my_model.yaml
# Created recipe: my_model.yaml
# Based on template: hybrid_resnet_se

info

Show information about a model/recipe.

tvc info --yaml model.yaml [--verbose]

Options:

--yaml, -y: Recipe YAML file (required)
--verbose, -v: Show detailed model information

Example:

tvc info --yaml model.yaml --verbose
# Recipe: model.yaml
# Name: ResNet50-SE-Custom
# Backbone: resnet50
#   Weights: IMAGENET1K_V2
#   Patches: 2 modifications
#
# Model Statistics:
#   Total parameters: 25,557,032
#   Trainable parameters: 2,048,100

YAML Recipe Format

Basic Structure

# Metadata
name: My Model
version: "1.0.0"

# Option 1: Hybrid backbone
backbone:
  name: resnet50
  weights: IMAGENET1K_V2
  patches:
    layer3: {wrap: se}

# Option 2: From scratch
stem: {type: conv, channels: 64}
stages:
  - {pattern: residual, channels: 64, blocks: 2}
  - {pattern: residual, channels: 128, blocks: 2, downsample: true}

# Head configuration
head:
  num_classes: 100
  dropout: 0.3

Macros

macros:
  attention: se
  dropout: 0.3

backbone:
  name: resnet50
  patches:
    layer3: {wrap: "@attention"}  # Expands to "se"

head:
  dropout: "@dropout"  # Expands to 0.3

Inheritance

# Extend a built-in template
extends: resnet_base

# Override specific parts
head:
  num_classes: 10